- Rufus Grig
Hello and welcome to The Learning Kerv, the podcast from Kerv that delves into the latest developments in information technology and explores how organisations can put them to work for the good of their people, their customers, society and the planet. My name's Rufus Grigg and in this series, with the help of some very special guests, we're going to be looking into all things generative AI. And this week I'm delighted to let you know that our special guest is none other than William Dorrington, CTO of Kerv's digital transformation practice, Kerv Digital, Microsoft Most Valued Professional and general all-round genius in the world of business technology near AI. Hi, Will, how are you doing?
- Will Dorrington
I'm doing very well. I mean, I feel I'm a bit set up for failure here, Rufus, but I'm excited to be on this. We've tried getting on this podcast a few times now and it's great we finally made it happen. So yeah, fantastic to be here. Thanks for having me.
- Rufus Grig
Not at all. It's brilliant. And I'm glad we've tracked you down on your holidays in Bratislava to be with us today. So unless you've been living under a rock for the last 18 months or so, Generative AI has had an impact on almost everybody, certainly in the workplace, in every aspect of business technology. I'm sure you'll have played with ChatGPT or Gemini or BARD or one of the other pieces out there. And while we are going to spend most of this series talking about Generative AI, we are going to spend the first episode a bit before Gen AI talking about the underlying technology, the basics of artificial intelligence itself. So that hopefully that gives us a bit of a grounding to be able to understand those core concepts and some of the opportunities and some of potentially the dangers and pitfalls that can come. So Will, perhaps you could start by answering, what is AI?
- Will Dorrington
Absolutely. Yeah, I think that's a really good place to kick this off. And throughout all the discussions we've had, we appreciate there's a wide, varied audience, or we hope there's a wide, varied audience. It might just be us listening back to this, Rufus. So I'm going to try and keep it as succinct and direct and high level as I can. So when I think of AI, or let's break it down even further, artificial intelligence, it's the ability of machines like computers to mimic human intelligence. I mean, that's its most rawest. So this can include things like learning from experience, understanding language, solving problems, even making decisions. Essentially. is about making machines think and act in a smart way like humans do. Well, most humans. I don't know if I always act as a smart way. And, you know, there's almost like an Arthur C. Clarke type quote here around it's sometimes indistinguishable from magic. When you really see AI at work, when you start seeing classification and identification, you go, wow, that is truly impressive.
- Rufus Grig
I think that Arthur C. Clarke phrase is really interesting because the magic is sort of ever present, isn't it? I mean, I remember when I first heard speech recognition. by a computer, it seemed extraordinary. And now the first thing I say every morning is Alexa, play Radio 4. You know, I talk to machines all day, every day, and it's completely commonplace.
- Will Dorrington
And it is, and it's around us in everyday life everywhere. So, I mean, I'm a big fan of Alexa. I use it for the most basic things like Alexa, what's the time, all the way to, okay, let's turn the downstairs lights on, let's turn the upstairs lights on, you know. We spoke about this before, didn't we? And spoke about sort of Windows Hello, which uses AI to recognize your face and log you in securely. I mean, there's some real lessons learned there. You know, we'll get on to how these models are trained shortly, but they didn't have a diverse enough training set and it caused them some issues. I mean, we won't go into it now, but it's definitely something that the audience members may want to look up. And even, I know, Netflix was an example we've spoke about before, how it suggests movies we may like and shows we may like based on what we've seen before. And actually, Netflix goes further than that. If they're thinking of investing in a new show, a new movie, they will actually use an algorithm to see whether they should invest in it and they will base decisions on the outcome of that. What's the likelihood of this actually being adopted and viewed by its audience based on what they're watching currently on a mass level? And then we could talk about BBC subtitles, we could talk about noise suppression on Teams calls and hopefully on this call, but it's constantly working behind the scenes to improve experiences across the board. There's been some real successes there and there's been great failures as well.
- Rufus Grig
The failures are always fun, aren't they? Okay, so it's not new, AI. It's something that we're all using, or most of us are using, all day, every day in our lives. And I know that the headlines, if you just read newspapers, you'd think it was invented 18 months ago with generative AI, but actually it's a fundamental part of all of our lives.
- Will Dorrington
Absolutely.
- Rufus Grig
So if we just break down the basics, because it's quite an umbrella term, machine learning is a term that's often coined. From my understanding, everything really is springing from this concept of machine learning, of the machines being able to learn from the environment around them. Could you just tell us a bit about what machine learning is?
- Will Dorrington
Yeah, sure. So if we compare it against traditional compute, so we look at machine learning and how it's different from traditional computing. Instead of actually being explicitly programmed with rules, machine learning allows a computer to actually learn from the data that it's reviewing. So in traditional computing, you give a computer a set of instructions, a logic pattern that it follows every time. It can be mainly linear, but there would be arguments of unit tranches and conditional branches there as well. But with machine learning, you provide the computer with the data and it will start figuring out patterns and make decisions on its own. So this means machine learning systems can improve and adapt over time, whereas traditional compute systems stay the same way unless we go in and reprogram it. There's ways of looking at this. So if machine learning is like making predictions without knowing all the rules up front. So imagine you're playing a new game. You go to a village funfair, village fates in England, and they always have those funny games which like guess how many sweets are in the jar and you get to win all these sweets and you can go home and eat them all and feel entirely sick. But actually that can be really hard to start giving random guesses of how many jelly beans are in a jar but if you start seeing the actual number of jelly beans in similar jars you can start to notice patterns and even if you don't know the exact method to count the jelly beans just by looking you can start getting better at predicting actually this is probably how much is actually in there and that's just before you even start counting them out so you can get better at predicting based on previous knowledge. Machine learning, the computer is doing something entirely similar. It's finding the patterns in data and it's making those predictions, even if it doesn't understand the underlying rules like a human might. And I know you've got one that's probably a lot less convoluted on school experiments.
- Rufus Grig
Yeah, well, I guess, yes, the scientist in me, I studied physics. I didn't study jelly beans in jars at Village Fates. But yeah, I guess you do an experiment, you take lots of measurements, you don't know what the rules are. I suppose you're doing an experiment to try and work out the rules, you plot a graph and you can find, you know, if I drop this thing from three meters high versus five meters high, how long does it take till it hits the ground? And I suppose it's that sort of similar, I'm collecting that sort of experimental data and then I'm spotting that pattern. In me and my physics lab, the pattern was a graph in machine learning, I guess. it's stored somewhere else and the machine's able to infer what's going on.
- Will Dorrington
Yeah, no, and I love that. I think it's such a really nice, simple way of explaining it. Maybe I should stop just ranting about jelly beans to people and maybe pick that one up instead and take full credit for it.
- Rufus Grig
By all means. Give us some typical machine learning applications. Where might it be a sensible approach to take to solve a business or a commercial problem?
- Will Dorrington
So let's look at several common applications of machine learning that we could encounter every day. So let's break that down to like prediction, clustering. classification and anomaly detection. I think they're the key ones. So although we're not sponsored by Netflix, let's do another Netflix sort of pattern here. So if we look at machine learning, it's used for prediction. So we've stated in Netflix, it will suggest the next film or channels you may like based on your viewing history. The same can be with membership renewals and propensity modeling. It will look at your likelihood of renewing a subscription based on your interaction that subscription, but also based on other people like you. and their renewal history over time. So prediction and propensity is two different, very classic ways of using machine learning. And another one that most people are aware of is classification. So when we look at this, this can be things like machine learning help outlook identify if an email coming in is spam or not. Is it another recruiter trying to sell services? But at the same time, classification can be, is this image a hot dog or a porcupine? And it can start navigating and classifying that for us. Then we get to clustering and anomaly detection. So this is where we can start grouping similar items together, items of data, and spotting if something is unusual. So this is what literal banks do. So you know, you look at my banking transaction at the moment, I've been to a lot of festivals, it's probably going beer, beer, hot dog, burger, couple glasses of rum. And as soon as I purchase a water, my bank goes, wait a minute, that's not like Will's purchasing history. There's something weird here. It goes outside the normal cluster. There's an anomaly. Let's block this card. Ring up Mr. Dorrington and say, are you sure you wanted to purchase this water? Because, you know, based on our prediction, based on our machine learning pattern, that should have been a rum and coke. And this is all ways that machine learning is used in our day-to-day environment, but it's using vast amounts of data when it applies it in these useful ways.
- Rufus Grig
Okay, so prediction, classification, and clustering, they're all effectively ways of helping a machine make a decision about something from some data. So how does a computer
- Will Dorrington
actually learn what is the process of building this up sure okay so let's break this down into a few stages so a computer learns in these key steps remember this this is high level this is any data scientist listening to this will start really sort of curling up and screaming right now but it often starts with a neural network which is a system inspired by the human brain hence the name neural network think of neurons and synapses etc all sparking away and that helps the computer to spot patterns I'm sure one day we'll do a podcast on a much deeper dive there or feed forward networks, et cetera, et cetera, but it will turn the audience off very quickly. So we've got the neural network, which is the ability to help the computer spot patterns. Let's keep it at that level. Then the computer is given labeled data. We'll talk about unlabeled data a bit later. So this is examples of data where we already know the right answer. So say we were feeding images. I don't know why I went with porcupines and hot dogs. I found it as weird as you, but let's carry on with that. we feed it a load of hot dog images and we say this is a hot dog these are different types of hot dogs these are different types of porcupines yes the porcupine may be angry but it's still a porcupine so it's all this labeled data it's getting to know the right answers to already we're telling it the right answers then we have model training so we've done neural network learning spot patterns labeled data feeding it the right answers now we get to model training and during model training the computer uses the data to adjust its approach to identifying in this case the classification approach to porcupine and the hot dog and it's getting better at making the predictions so in hope that once we've trained it we can then use some test data and get it to actually make those predictions so we give it the test data which are brand new examples that it has not seen before it's not labeled we're not saying this is an angry porcupine this is a new york star hot dog and we can see how well it has learned and then we can adjust if needed and if it's not accurate enough we will tweak, we will refine that model until it's performing as expected or as close to expected. And then over time, the computer can process this and get better at the task it's been designed to do. There's a lot more nuance to it, but that is the key part. So neural network, feeding labeled data, do a bit of model training, test it, get hopefully a successful outcome.
- Rufus Grig
Okay. So a neural network's almost like the blank mind. have some data that you understand, so you know the rules that align to that data, you feed it into the model, and then you have some data that you test that it's actually performing, and then you sort of keep going again until it's working well.
- Will Dorrington
That's absolutely it, until you get your favourable outcome. And it can be a bit frustrating, and actually making sure you pick the right pattern is quite hard, but that's why you have very clever data scientists and machine learning engineers, which I think we do discuss a bit later in the podcast.
- Rufus Grig
Yeah, okay. So I guess it's important that data is good. And I guess also that test data is going to be vital in being able to do that validation, because we're putting these systems into practice. I mean, we've talked about some quite fun examples, but I guess some others might be classifying skin blemishes or something in dermatology or classifying some medical symptoms, you know, if do these symptoms look like this outcome, and these systems need quite a lot of trust associated with them.
- Will Dorrington
No, indeed. And detecting faulty sort of cladding on buildings, you know, after the big Grenfell Tower disaster. I mean, there's been lots of these high profile uses of AI. So, yeah, it's not all porcupines and hot dogs. There are some genuine real applications of this as well. And, you know, data is everything in tech, even down to if we ignore AI and talk about applications and solutions and, you know, from Genesis solutions to dynamic solutions to, dare I say, Salesforce, et cetera. It's all just data. It's either processing data. transforming data and adding data in a validated way or deleting data but everything is about the movement of data in what we do in tech so I think it's probably best we actually dive into a bit of data because with machine learning it's no different data is the heart of machine learning and you can use all sorts of data for machine learning so from numbers and text to images to sound and even video so essentially if it can be stored and processed by a computer it can be used for machine learning So as for the process of actually collecting this, it usually starts by gathering all that delicious raw data from various sources. I mean, ChatGPT did some great stuff around actually grabbing everything from, you know, Wikipedia, journals, even places where you probably wouldn't want to grab data from like Reddit and open forums because, you know, let's face it, not everything on the Internet should be digested. And even things like IoT and sensors and grabbing all those event logs from user interactions. And then once you get all this and you think you've got your right. amount of data that you need for what you're training. You then need to cleanse and organize this data. You need to remove any errors, any duplicates, any irrelevant or useless information. And then once that is cleaned, once that's been often labeled, we might go for an unlabeled structure. By that, I mean sort of adding the correct answers and categories to each piece of data. We can then go on to the next step, which is actually preparing that data and splitting it into either training or testing sets so that the model can learn from this. and actually evaluate off of each other. So we can learn from the training set, and then on the testing set, it can be evaluated on a bit like what I was saying about the labeled images, then given an unlabeled image to see if it comes out. And this more structured approach to data, and it isn't always structured approach, is what allows machine learning to actually be trained really effectively. So you know, if you ever have spare time to the listeners, I would really look at what ChatGPT did there because it's entirely impressive. We're talking huge volumes of data.
- Rufus Grig
In thinking about the role of a data scientist, then their job is in understanding how do I put together data that's going to be effective for training the model? And I guess that is going to involve things like removing outliers or data that's got errors in it and making sure that we're not effectively feeding garbage in to the model where you presumably, I'm guessing, as in most computing paradigms, you then get garbage out.
- Will Dorrington
Absolutely. They're there to... to make sure they've got that right set of data, make sure that the data does have an element of patterns, trends, and that they can get those insights. So yeah, we always say garbage in, garbage out. Absolutely.
- Rufus Grig
Okay. We have used a bit of jargon so far in this conversation. I do want to just make sure we understand that. So model. Model and algorithm we've talked about. What is a model? What is an algorithm? Are they the same thing? Are they slightly different?
- Will Dorrington
The way I think of this, and it is a really simple way of explaining, and maybe too simple, but I think of an algorithm as a recipe. So it's like the step-by-step instructions the computer follows to actually learn from the data, where the model is what you get at the end. It's the new finished dish from the recipe. It's the trained result, which can be used to make the predictions or decisions based on the new data. So the algorithm is the recipe. The model is actually the dish that you get to eat at the end.
- Rufus Grig
Okay. So a model is something that we can consume. You've got a model that, I don't know, identifies different types of flours or can produce. predict how much a house is going to cost based on how many bedrooms it's got and where it is. And it's the computer scientists worry about the algorithms inside as consumers or as users, we think about the model and the way it's used.
- Will Dorrington
Spot on. Yeah. Yeah. That's how I see the world anyway.
- Rufus Grig
Okay. I understand that. You've talked about labelled data and unlabelled data. Just explain those for us.
- Will Dorrington
So labelled data is when each piece of data actually comes with the correct answer. So I'm sure afterwards you're going to say, Will, please don't talk about hot dogs and porcupines on the next podcast that we invite you back. But this is where we say to the algorithm, you know, and to why we trained the model. This is a hot dog. This is a porcupine. You know, we could do that with cats and dogs and, you know, and actually classifying spam and not spam emails and lots of other stuff that we'd use. Where unlabeled data does not have these answers. The computer has to figure out on its own. It starts grouping. pieces of data that he thinks are the same and those are outliers etc etc so labeled data we explicitly say this is what it is unlabeled data if we go figure out for yourself good luck we've talked about supervised and unsupervised learning as well what are those okay so yeah i wrote a blog on this recently because this comes up a lot and the best way i can explain this is supervised learning is like being a student with a teacher who provides the correct answers during training it's a bit like You're playing Monopoly, but you've never played it before. And you're sitting down with your friends and family, although playing Monopoly with families is occasionally a no-no, as we know all the jokes are flipping.
- Rufus Grig
War zone in my family, if you play Monopoly. Absolute war zone.
- Will Dorrington
And they're teaching you how to play it. They're saying, here's the tokens. Here's how you roll the dice. Here's how you collect the houses that ran. Here's how you have a big fight. Where unsupervised learning is more about exploring on your own, where the computer finds the patterns and relationships without any guidance. So you're just sitting there. watching your friends and family play, but they're not telling you how to play it. They're like, just watch and learn. You can just pick it up. Okay. That is the difference. So supervised, you're explicitly being provided the approach and the correct answers during training where unsupervised learning, you're not being provided that the computer figures it out on its own.
- Rufus Grig
Great. Getting there. Training. You've talked about training the model a lot. We've not used the word, but I hear it talked a lot about inference. So what is training? What is inference? And how should we think about those in the context of machine learning?
- Will Dorrington
So the training phase is where the computer learns from the data. So it's improving the model where the inference is, where the trained model is put to use. So it's making the predictions or the decisions based on new or unseen data. So training is improving the model. Inference is actually putting it to use. I think that's the most simple way of explaining. I mean, Rufus, I don't know if you've got any comment on that at all.
- Rufus Grig
I guess you hear about people talking about the training and inference phases of machine learning, and particularly, I suppose, in terms of the speed. You know, it takes a long time to train a model because I'm feeding it lots of data. Inference is a much quicker process because I show it one instance of a record and it feeds it through, or I'm showing it one image and it's classifying or identifying it. And I guess also in terms of the power consumption, obviously training a model is a very hungry thing to do. It sucks up lots of compute resource, lots of power, whereas the inference is something that goes much more quickly. Right. Thank you. Well, we had a bit of a jargon bust there. I know we've talked about images and we talked about sound as well, but computers, you tend to think of as dealing with numbers. Am I right to think that the computer doesn't really know that it's an image? It's dealing with a bunch of numbers and looking at...
- Will Dorrington
patterns there is do we have to think differently about images and text and sound in this world yeah i remember when computer vision so let's focus on images because i think that it was something that struck me when computer vision first came out and things like ocr etc etc and it fascinated me i was thinking how how have they trained the computer to see and interpret images because it's obviously not just looking at it there's got to be something around machine learning so here's how i sort of explain how it works so when a computer processes an image it doesn't see the picture like we do instead it converts that into a grid of numbers because we know it's all computational mathematics right all the way back to Alan Turing's paper on computational mathematics that still is the way computers work to this day so if we think about on an image each number that is so it's a string of numbers it converts the image into a grid of numbers and each number represents a pixel so a tiny dot of color in the image and then its value will correspond to the color and the intensity of that pixel. So if it's transparent, if how opaque it is, and if we choose a black and white image, you know, I think it reduces the complexity. So these numbers will range from, I think zero is black and 255 is white. And if that's wrong, it doesn't matter. It's just an example. And in between that zero and 255, there'll be a different shades of gray. So what it does, it starts converting these into numbers. Every pixel becomes a number and it can start picking up on all these different. patterns so the machine learning can start analyzing the patterns to recognize objects faces or even actually detect anomalies because that's how it sees it's used to seeing that way so the model is trained often on thousands or even at times you can actually get millions and millions of labeled images you see some of the stuff amazon and facebook do and it can be photos like cats and dogs or porcupines and hot dogs as we all love on this show and then it can learn to identify accurately new images now there's something some really brilliant disasters around this that we won't go into we're going to try to keep this positive but it is feeding through those images converting to those gridded numbers understanding then the colors the intensity that you can start training and identifying what it's seeing so it is a form of machine learning and it is the computers being trained to understand and interpret those images by breaking them down into numbers and patterns it's the most simplest way of explaining it okay
- Rufus Grig
no i get that you And I also know that your first novel is now going to be 255 Shades of Grey. And so I suppose it's effectively, it's breaking those letters and number images on a page into that series of pixels, which you can then recognize those patterns and turn them into words. And then we get all those amazing applications such as you can now hold your smartphone up and it will translate the menu in your French restaurant into English.
- Will Dorrington
sort of assistance for visually impaired people to be able to read that's brilliant springing from that that optical character recognition all driven by effectively batches of numbers absolutely and uh you know i think uh at a later date we'll maybe talk about diffusion models as well which is you know the other way around where it's actually creating images by uh interpreting what we've asked so what we've seen with things like dali etc i mean because that's that's a really interesting topic in itself we'll get into that in the next episode definitely okay so i think i'm getting a great understanding we've
- Rufus Grig
We collect data. We've trained models with data. We can then infer by putting our test samples through those models and it can either predict what it might be. It can classify it. It can detect anomalies and all where we don't actually necessarily know the hard and fast rules that set things out.
- Will Dorrington
Absolutely.
- Rufus Grig
Really useful. Thank you. I mentioned data science earlier. Is there such a, is there a categorization of what is a data scientist? And if so, What do they do? What's a data scientist's role specifically?
- Will Dorrington
So it can get quite complex, especially when you look at, you know, the key data areas of data engineering, data science, data visualization. And when you go into data science and data engineering, there's a gap between that for machine learning engineering. There's different parts of visualization. So it does get very... convoluted and it breaks down and down and down and half the time when people say they want a data science they actually want a data engineer so i think it's a really good question and i think a good way to define a data science is someone who uses data to solve complex problems that's at its most basic but the key thing is they actually analyze large sets of data looking for patterns trends and insights that can help make better decisions or predictions so data scientists use a mix of skills that include statistics so these are actuarial mathematicians you know i'm fortunate enough to have a few really strong data scientist friends and they will have doctorates in actuarial mathematics etc incredibly impressive human beings but not only do they have a really vast and deep complex understanding of maths they're also brilliant programmers they know how to use programming languages and they already use python etc to create these models and then they also have a ton of domain knowledge depending on the area they've chose to specialize in you know actually one of my friends did housing for a long time and now he works over at amazon So we could turn that raw data and data sciences can turn that raw data into incredibly valuable information because not only can they start applying incredibly advanced statistical models and mathematical models, they can program absolute machine learning models over the top and then apply domain knowledge to really get the best out of it. A bit like where we're very industry aligned at Kerv because we know domain knowledge is very key. But on the other hand, and let's do look at the flip side of this, where a machine learning engineer is actually more focused on building and deploying the machine learning models. So they'll actually deploy a lot of the research that's already gone in. They'll take the insights and models developed by the data scientists and make them into real world applications. While both are involved with working with data and machine learning, data science are often more focused on the exploration and the analysis. Machine learning engineers much more on the practical implementation and optimization of those models.
- Rufus Grig
OK, fascinating. I mean, it's an amazing mix of skills that you. understand the maths and the statistics you understand data you know how to program the computer to be able to handle it and you understand the subject matter and the domain in which you're working at these these must be fairly rare beasts they are i'd say they're probably the hardest at the moment on the market to find i mean we know that we've got some phenomenal
- Will Dorrington
talent inside Kerv but data scientists right now are to find a really good one are yeah very dusty it's it's tough and just the sort of
- Rufus Grig
tools that they use. I mean, you mentioned Python, which is a language, and I'm guessing cloud computing is used because of the sort of the size of the data sets that we're using. Is that the sort of tooling that these people are using?
- Will Dorrington
Absolutely. So it's the ability to build those statistical models. A lot of them are cloud-based these days. Actually understanding of even various databases to use, vector databases, and vector databases, of course, are very important when it comes to building these machine learning models. and actually just having an appreciation of also applications and ecosystem architecture in general. Whenever I speak to my sort of data sciences friends, I'm impressed by how much they know outside that space because they want to know about how to get good quality data in. And that is by CRM systems, ERP systems, portals. But it all comes back down to those statistical tools that they can use, you know, Python, et cetera, et cetera.
- Rufus Grig
Thanks, Will. It's been really interesting understanding that sort of background of AI and machine learning. Just give us a sneak peek of what we're going to be talking about next time.
- Will Dorrington
OK, so I think we've got a really good, hopefully got a really good understanding of some of the basics and fundamentals of AI and machine learning. You know, we've explored applications. We've explored how we train these models, how we use data. We've busted some of that jargon as well. We've even spoke about the roles involved. I think now we're at a perfect place to actually jump into the next stage of AI, the hype that has occurred around, you know, chat GPT, around generative pre-trained transformer architecture, the bi-directional encoder models from Google, and large language models all up, and the acceleration we've seen of just taking text, inputting text, and having new and creative text at the back of that, to then images, to sound, to videos, and I think that's where we're going to move to next. this next stage of AI, the AI era we're now in, which has been powered by generative AI.
- Rufus Grig
Thank you, Will. That has been fascinating. I can't wait to get into that conversation. If you've been interested in what we've had to say, then please do get in touch and tell us what you think. You can find out more about Kerv and about AI and about Will in general by visiting Kerv.com and do listen out for that next episode. You can subscribe, you can tell all your friends We really look forward to speaking to you. Thank you so much, Will. William Dorrington, CTO of Kerv Digital. Really look forward to getting stuck into the generative AI piece next time. In the meantime, thank you for listening. Will Barron
